1 /* Copyright (C) 2009 Red Hat, Inc.
2 * Copyright (C) 2006 Rusty Russell IBM Corporation
4 * Author: Michael S. Tsirkin <mst@redhat.com>
6 * Inspiration, some code, and most witty comments come from
7 * Documentation/virtual/lguest/lguest.c, by Rusty Russell
9 * This work is licensed under the terms of the GNU GPL, version 2.
11 * Generic code for virtio server in host kernel.
14 #include <linux/eventfd.h>
15 #include <linux/vhost.h>
16 #include <linux/uio.h>
18 #include <linux/mmu_context.h>
19 #include <linux/miscdevice.h>
20 #include <linux/mutex.h>
21 #include <linux/poll.h>
22 #include <linux/file.h>
23 #include <linux/highmem.h>
24 #include <linux/slab.h>
25 #include <linux/vmalloc.h>
26 #include <linux/kthread.h>
27 #include <linux/cgroup.h>
28 #include <linux/module.h>
29 #include <linux/sort.h>
30 #include <linux/sched/mm.h>
31 #include <linux/sched/signal.h>
32 #include <linux/interval_tree_generic.h>
36 static ushort max_mem_regions = 64;
37 module_param(max_mem_regions, ushort, 0444);
38 MODULE_PARM_DESC(max_mem_regions,
39 "Maximum number of memory regions in memory map. (default: 64)");
40 static int max_iotlb_entries = 2048;
41 module_param(max_iotlb_entries, int, 0444);
42 MODULE_PARM_DESC(max_iotlb_entries,
43 "Maximum number of iotlb entries. (default: 2048)");
46 VHOST_MEMORY_F_LOG = 0x1,
49 #define vhost_used_event(vq) ((__virtio16 __user *)&vq->avail->ring[vq->num])
50 #define vhost_avail_event(vq) ((__virtio16 __user *)&vq->used->ring[vq->num])
52 INTERVAL_TREE_DEFINE(struct vhost_umem_node,
53 rb, __u64, __subtree_last,
54 START, LAST, static inline, vhost_umem_interval_tree);
56 #ifdef CONFIG_VHOST_CROSS_ENDIAN_LEGACY
57 static void vhost_disable_cross_endian(struct vhost_virtqueue *vq)
59 vq->user_be = !virtio_legacy_is_little_endian();
62 static void vhost_enable_cross_endian_big(struct vhost_virtqueue *vq)
67 static void vhost_enable_cross_endian_little(struct vhost_virtqueue *vq)
72 static long vhost_set_vring_endian(struct vhost_virtqueue *vq, int __user *argp)
74 struct vhost_vring_state s;
79 if (copy_from_user(&s, argp, sizeof(s)))
82 if (s.num != VHOST_VRING_LITTLE_ENDIAN &&
83 s.num != VHOST_VRING_BIG_ENDIAN)
86 if (s.num == VHOST_VRING_BIG_ENDIAN)
87 vhost_enable_cross_endian_big(vq);
89 vhost_enable_cross_endian_little(vq);
94 static long vhost_get_vring_endian(struct vhost_virtqueue *vq, u32 idx,
97 struct vhost_vring_state s = {
102 if (copy_to_user(argp, &s, sizeof(s)))
108 static void vhost_init_is_le(struct vhost_virtqueue *vq)
110 /* Note for legacy virtio: user_be is initialized at reset time
111 * according to the host endianness. If userspace does not set an
112 * explicit endianness, the default behavior is native endian, as
113 * expected by legacy virtio.
115 vq->is_le = vhost_has_feature(vq, VIRTIO_F_VERSION_1) || !vq->user_be;
118 static void vhost_disable_cross_endian(struct vhost_virtqueue *vq)
122 static long vhost_set_vring_endian(struct vhost_virtqueue *vq, int __user *argp)
127 static long vhost_get_vring_endian(struct vhost_virtqueue *vq, u32 idx,
133 static void vhost_init_is_le(struct vhost_virtqueue *vq)
135 vq->is_le = vhost_has_feature(vq, VIRTIO_F_VERSION_1)
136 || virtio_legacy_is_little_endian();
138 #endif /* CONFIG_VHOST_CROSS_ENDIAN_LEGACY */
140 static void vhost_reset_is_le(struct vhost_virtqueue *vq)
142 vhost_init_is_le(vq);
145 struct vhost_flush_struct {
146 struct vhost_work work;
147 struct completion wait_event;
150 static void vhost_flush_work(struct vhost_work *work)
152 struct vhost_flush_struct *s;
154 s = container_of(work, struct vhost_flush_struct, work);
155 complete(&s->wait_event);
158 static void vhost_poll_func(struct file *file, wait_queue_head_t *wqh,
161 struct vhost_poll *poll;
163 poll = container_of(pt, struct vhost_poll, table);
165 add_wait_queue(wqh, &poll->wait);
168 static int vhost_poll_wakeup(wait_queue_entry_t *wait, unsigned mode, int sync,
171 struct vhost_poll *poll = container_of(wait, struct vhost_poll, wait);
173 if (!(key_to_poll(key) & poll->mask))
176 vhost_poll_queue(poll);
180 void vhost_work_init(struct vhost_work *work, vhost_work_fn_t fn)
182 clear_bit(VHOST_WORK_QUEUED, &work->flags);
184 init_waitqueue_head(&work->done);
186 EXPORT_SYMBOL_GPL(vhost_work_init);
188 /* Init poll structure */
189 void vhost_poll_init(struct vhost_poll *poll, vhost_work_fn_t fn,
190 __poll_t mask, struct vhost_dev *dev)
192 init_waitqueue_func_entry(&poll->wait, vhost_poll_wakeup);
193 init_poll_funcptr(&poll->table, vhost_poll_func);
198 vhost_work_init(&poll->work, fn);
200 EXPORT_SYMBOL_GPL(vhost_poll_init);
202 /* Start polling a file. We add ourselves to file's wait queue. The caller must
203 * keep a reference to a file until after vhost_poll_stop is called. */
204 int vhost_poll_start(struct vhost_poll *poll, struct file *file)
212 mask = file->f_op->poll(file, &poll->table);
214 vhost_poll_wakeup(&poll->wait, 0, 0, poll_to_key(mask));
215 if (mask & POLLERR) {
217 remove_wait_queue(poll->wqh, &poll->wait);
223 EXPORT_SYMBOL_GPL(vhost_poll_start);
225 /* Stop polling a file. After this function returns, it becomes safe to drop the
226 * file reference. You must also flush afterwards. */
227 void vhost_poll_stop(struct vhost_poll *poll)
230 remove_wait_queue(poll->wqh, &poll->wait);
234 EXPORT_SYMBOL_GPL(vhost_poll_stop);
236 void vhost_work_flush(struct vhost_dev *dev, struct vhost_work *work)
238 struct vhost_flush_struct flush;
241 init_completion(&flush.wait_event);
242 vhost_work_init(&flush.work, vhost_flush_work);
244 vhost_work_queue(dev, &flush.work);
245 wait_for_completion(&flush.wait_event);
248 EXPORT_SYMBOL_GPL(vhost_work_flush);
250 /* Flush any work that has been scheduled. When calling this, don't hold any
251 * locks that are also used by the callback. */
252 void vhost_poll_flush(struct vhost_poll *poll)
254 vhost_work_flush(poll->dev, &poll->work);
256 EXPORT_SYMBOL_GPL(vhost_poll_flush);
258 void vhost_work_queue(struct vhost_dev *dev, struct vhost_work *work)
263 if (!test_and_set_bit(VHOST_WORK_QUEUED, &work->flags)) {
264 /* We can only add the work to the list after we're
265 * sure it was not in the list.
266 * test_and_set_bit() implies a memory barrier.
268 llist_add(&work->node, &dev->work_list);
269 wake_up_process(dev->worker);
272 EXPORT_SYMBOL_GPL(vhost_work_queue);
274 /* A lockless hint for busy polling code to exit the loop */
275 bool vhost_has_work(struct vhost_dev *dev)
277 return !llist_empty(&dev->work_list);
279 EXPORT_SYMBOL_GPL(vhost_has_work);
281 void vhost_poll_queue(struct vhost_poll *poll)
283 vhost_work_queue(poll->dev, &poll->work);
285 EXPORT_SYMBOL_GPL(vhost_poll_queue);
287 static void __vhost_vq_meta_reset(struct vhost_virtqueue *vq)
291 for (j = 0; j < VHOST_NUM_ADDRS; j++)
292 vq->meta_iotlb[j] = NULL;
295 static void vhost_vq_meta_reset(struct vhost_dev *d)
299 for (i = 0; i < d->nvqs; ++i)
300 __vhost_vq_meta_reset(d->vqs[i]);
303 static void vhost_vq_reset(struct vhost_dev *dev,
304 struct vhost_virtqueue *vq)
310 vq->last_avail_idx = 0;
312 vq->last_used_idx = 0;
313 vq->signalled_used = 0;
314 vq->signalled_used_valid = false;
316 vq->log_used = false;
317 vq->log_addr = -1ull;
318 vq->private_data = NULL;
319 vq->acked_features = 0;
321 vq->error_ctx = NULL;
327 vhost_reset_is_le(vq);
328 vhost_disable_cross_endian(vq);
329 vq->busyloop_timeout = 0;
332 __vhost_vq_meta_reset(vq);
335 static int vhost_worker(void *data)
337 struct vhost_dev *dev = data;
338 struct vhost_work *work, *work_next;
339 struct llist_node *node;
340 mm_segment_t oldfs = get_fs();
346 /* mb paired w/ kthread_stop */
347 set_current_state(TASK_INTERRUPTIBLE);
349 if (kthread_should_stop()) {
350 __set_current_state(TASK_RUNNING);
354 node = llist_del_all(&dev->work_list);
358 node = llist_reverse_order(node);
359 /* make sure flag is seen after deletion */
361 llist_for_each_entry_safe(work, work_next, node, node) {
362 clear_bit(VHOST_WORK_QUEUED, &work->flags);
363 __set_current_state(TASK_RUNNING);
374 static void vhost_vq_free_iovecs(struct vhost_virtqueue *vq)
384 /* Helper to allocate iovec buffers for all vqs. */
385 static long vhost_dev_alloc_iovecs(struct vhost_dev *dev)
387 struct vhost_virtqueue *vq;
390 for (i = 0; i < dev->nvqs; ++i) {
392 vq->indirect = kmalloc(sizeof *vq->indirect * UIO_MAXIOV,
394 vq->log = kmalloc(sizeof *vq->log * UIO_MAXIOV, GFP_KERNEL);
395 vq->heads = kmalloc(sizeof *vq->heads * UIO_MAXIOV, GFP_KERNEL);
396 if (!vq->indirect || !vq->log || !vq->heads)
403 vhost_vq_free_iovecs(dev->vqs[i]);
407 static void vhost_dev_free_iovecs(struct vhost_dev *dev)
411 for (i = 0; i < dev->nvqs; ++i)
412 vhost_vq_free_iovecs(dev->vqs[i]);
415 void vhost_dev_init(struct vhost_dev *dev,
416 struct vhost_virtqueue **vqs, int nvqs)
418 struct vhost_virtqueue *vq;
423 mutex_init(&dev->mutex);
425 dev->log_file = NULL;
430 init_llist_head(&dev->work_list);
431 init_waitqueue_head(&dev->wait);
432 INIT_LIST_HEAD(&dev->read_list);
433 INIT_LIST_HEAD(&dev->pending_list);
434 spin_lock_init(&dev->iotlb_lock);
437 for (i = 0; i < dev->nvqs; ++i) {
443 mutex_init(&vq->mutex);
444 vhost_vq_reset(dev, vq);
446 vhost_poll_init(&vq->poll, vq->handle_kick,
450 EXPORT_SYMBOL_GPL(vhost_dev_init);
452 /* Caller should have device mutex */
453 long vhost_dev_check_owner(struct vhost_dev *dev)
455 /* Are you the owner? If not, I don't think you mean to do that */
456 return dev->mm == current->mm ? 0 : -EPERM;
458 EXPORT_SYMBOL_GPL(vhost_dev_check_owner);
460 struct vhost_attach_cgroups_struct {
461 struct vhost_work work;
462 struct task_struct *owner;
466 static void vhost_attach_cgroups_work(struct vhost_work *work)
468 struct vhost_attach_cgroups_struct *s;
470 s = container_of(work, struct vhost_attach_cgroups_struct, work);
471 s->ret = cgroup_attach_task_all(s->owner, current);
474 static int vhost_attach_cgroups(struct vhost_dev *dev)
476 struct vhost_attach_cgroups_struct attach;
478 attach.owner = current;
479 vhost_work_init(&attach.work, vhost_attach_cgroups_work);
480 vhost_work_queue(dev, &attach.work);
481 vhost_work_flush(dev, &attach.work);
485 /* Caller should have device mutex */
486 bool vhost_dev_has_owner(struct vhost_dev *dev)
490 EXPORT_SYMBOL_GPL(vhost_dev_has_owner);
492 /* Caller should have device mutex */
493 long vhost_dev_set_owner(struct vhost_dev *dev)
495 struct task_struct *worker;
498 /* Is there an owner already? */
499 if (vhost_dev_has_owner(dev)) {
504 /* No owner, become one */
505 dev->mm = get_task_mm(current);
506 worker = kthread_create(vhost_worker, dev, "vhost-%d", current->pid);
507 if (IS_ERR(worker)) {
508 err = PTR_ERR(worker);
512 dev->worker = worker;
513 wake_up_process(worker); /* avoid contributing to loadavg */
515 err = vhost_attach_cgroups(dev);
519 err = vhost_dev_alloc_iovecs(dev);
525 kthread_stop(worker);
534 EXPORT_SYMBOL_GPL(vhost_dev_set_owner);
536 struct vhost_umem *vhost_dev_reset_owner_prepare(void)
538 return kvzalloc(sizeof(struct vhost_umem), GFP_KERNEL);
540 EXPORT_SYMBOL_GPL(vhost_dev_reset_owner_prepare);
542 /* Caller should have device mutex */
543 void vhost_dev_reset_owner(struct vhost_dev *dev, struct vhost_umem *umem)
547 vhost_dev_cleanup(dev, true);
549 /* Restore memory to default empty mapping. */
550 INIT_LIST_HEAD(&umem->umem_list);
552 /* We don't need VQ locks below since vhost_dev_cleanup makes sure
553 * VQs aren't running.
555 for (i = 0; i < dev->nvqs; ++i)
556 dev->vqs[i]->umem = umem;
558 EXPORT_SYMBOL_GPL(vhost_dev_reset_owner);
560 void vhost_dev_stop(struct vhost_dev *dev)
564 for (i = 0; i < dev->nvqs; ++i) {
565 if (dev->vqs[i]->kick && dev->vqs[i]->handle_kick) {
566 vhost_poll_stop(&dev->vqs[i]->poll);
567 vhost_poll_flush(&dev->vqs[i]->poll);
571 EXPORT_SYMBOL_GPL(vhost_dev_stop);
573 static void vhost_umem_free(struct vhost_umem *umem,
574 struct vhost_umem_node *node)
576 vhost_umem_interval_tree_remove(node, &umem->umem_tree);
577 list_del(&node->link);
582 static void vhost_umem_clean(struct vhost_umem *umem)
584 struct vhost_umem_node *node, *tmp;
589 list_for_each_entry_safe(node, tmp, &umem->umem_list, link)
590 vhost_umem_free(umem, node);
595 static void vhost_clear_msg(struct vhost_dev *dev)
597 struct vhost_msg_node *node, *n;
599 spin_lock(&dev->iotlb_lock);
601 list_for_each_entry_safe(node, n, &dev->read_list, node) {
602 list_del(&node->node);
606 list_for_each_entry_safe(node, n, &dev->pending_list, node) {
607 list_del(&node->node);
611 spin_unlock(&dev->iotlb_lock);
614 /* Caller should have device mutex if and only if locked is set */
615 void vhost_dev_cleanup(struct vhost_dev *dev, bool locked)
619 for (i = 0; i < dev->nvqs; ++i) {
620 if (dev->vqs[i]->error_ctx)
621 eventfd_ctx_put(dev->vqs[i]->error_ctx);
622 if (dev->vqs[i]->error)
623 fput(dev->vqs[i]->error);
624 if (dev->vqs[i]->kick)
625 fput(dev->vqs[i]->kick);
626 if (dev->vqs[i]->call_ctx)
627 eventfd_ctx_put(dev->vqs[i]->call_ctx);
628 if (dev->vqs[i]->call)
629 fput(dev->vqs[i]->call);
630 vhost_vq_reset(dev, dev->vqs[i]);
632 vhost_dev_free_iovecs(dev);
634 eventfd_ctx_put(dev->log_ctx);
638 dev->log_file = NULL;
639 /* No one will access memory at this point */
640 vhost_umem_clean(dev->umem);
642 vhost_umem_clean(dev->iotlb);
644 vhost_clear_msg(dev);
645 wake_up_interruptible_poll(&dev->wait, POLLIN | POLLRDNORM);
646 WARN_ON(!llist_empty(&dev->work_list));
648 kthread_stop(dev->worker);
655 EXPORT_SYMBOL_GPL(vhost_dev_cleanup);
657 static int log_access_ok(void __user *log_base, u64 addr, unsigned long sz)
659 u64 a = addr / VHOST_PAGE_SIZE / 8;
661 /* Make sure 64 bit math will not overflow. */
662 if (a > ULONG_MAX - (unsigned long)log_base ||
663 a + (unsigned long)log_base > ULONG_MAX)
666 return access_ok(VERIFY_WRITE, log_base + a,
667 (sz + VHOST_PAGE_SIZE * 8 - 1) / VHOST_PAGE_SIZE / 8);
670 static bool vhost_overflow(u64 uaddr, u64 size)
672 /* Make sure 64 bit math will not overflow. */
673 return uaddr > ULONG_MAX || size > ULONG_MAX || uaddr > ULONG_MAX - size;
676 /* Caller should have vq mutex and device mutex. */
677 static int vq_memory_access_ok(void __user *log_base, struct vhost_umem *umem,
680 struct vhost_umem_node *node;
685 list_for_each_entry(node, &umem->umem_list, link) {
686 unsigned long a = node->userspace_addr;
688 if (vhost_overflow(node->userspace_addr, node->size))
692 if (!access_ok(VERIFY_WRITE, (void __user *)a,
695 else if (log_all && !log_access_ok(log_base,
703 static inline void __user *vhost_vq_meta_fetch(struct vhost_virtqueue *vq,
704 u64 addr, unsigned int size,
707 const struct vhost_umem_node *node = vq->meta_iotlb[type];
712 return (void *)(uintptr_t)(node->userspace_addr + addr - node->start);
715 /* Can we switch to this memory table? */
716 /* Caller should have device mutex but not vq mutex */
717 static int memory_access_ok(struct vhost_dev *d, struct vhost_umem *umem,
722 for (i = 0; i < d->nvqs; ++i) {
726 mutex_lock(&d->vqs[i]->mutex);
727 log = log_all || vhost_has_feature(d->vqs[i], VHOST_F_LOG_ALL);
728 /* If ring is inactive, will check when it's enabled. */
729 if (d->vqs[i]->private_data)
730 ok = vq_memory_access_ok(d->vqs[i]->log_base,
734 mutex_unlock(&d->vqs[i]->mutex);
741 static int translate_desc(struct vhost_virtqueue *vq, u64 addr, u32 len,
742 struct iovec iov[], int iov_size, int access);
744 static int vhost_copy_to_user(struct vhost_virtqueue *vq, void __user *to,
745 const void *from, unsigned size)
750 return __copy_to_user(to, from, size);
752 /* This function should be called after iotlb
753 * prefetch, which means we're sure that all vq
754 * could be access through iotlb. So -EAGAIN should
755 * not happen in this case.
758 void __user *uaddr = vhost_vq_meta_fetch(vq,
759 (u64)(uintptr_t)to, size,
763 return __copy_to_user(uaddr, from, size);
765 ret = translate_desc(vq, (u64)(uintptr_t)to, size, vq->iotlb_iov,
766 ARRAY_SIZE(vq->iotlb_iov),
770 iov_iter_init(&t, WRITE, vq->iotlb_iov, ret, size);
771 ret = copy_to_iter(from, size, &t);
779 static int vhost_copy_from_user(struct vhost_virtqueue *vq, void *to,
780 void __user *from, unsigned size)
785 return __copy_from_user(to, from, size);
787 /* This function should be called after iotlb
788 * prefetch, which means we're sure that vq
789 * could be access through iotlb. So -EAGAIN should
790 * not happen in this case.
792 void __user *uaddr = vhost_vq_meta_fetch(vq,
793 (u64)(uintptr_t)from, size,
798 return __copy_from_user(to, uaddr, size);
800 ret = translate_desc(vq, (u64)(uintptr_t)from, size, vq->iotlb_iov,
801 ARRAY_SIZE(vq->iotlb_iov),
804 vq_err(vq, "IOTLB translation failure: uaddr "
805 "%p size 0x%llx\n", from,
806 (unsigned long long) size);
809 iov_iter_init(&f, READ, vq->iotlb_iov, ret, size);
810 ret = copy_from_iter(to, size, &f);
819 static void __user *__vhost_get_user_slow(struct vhost_virtqueue *vq,
820 void __user *addr, unsigned int size,
825 ret = translate_desc(vq, (u64)(uintptr_t)addr, size, vq->iotlb_iov,
826 ARRAY_SIZE(vq->iotlb_iov),
829 vq_err(vq, "IOTLB translation failure: uaddr "
830 "%p size 0x%llx\n", addr,
831 (unsigned long long) size);
835 if (ret != 1 || vq->iotlb_iov[0].iov_len != size) {
836 vq_err(vq, "Non atomic userspace memory access: uaddr "
837 "%p size 0x%llx\n", addr,
838 (unsigned long long) size);
842 return vq->iotlb_iov[0].iov_base;
845 /* This function should be called after iotlb
846 * prefetch, which means we're sure that vq
847 * could be access through iotlb. So -EAGAIN should
848 * not happen in this case.
850 static inline void __user *__vhost_get_user(struct vhost_virtqueue *vq,
851 void *addr, unsigned int size,
854 void __user *uaddr = vhost_vq_meta_fetch(vq,
855 (u64)(uintptr_t)addr, size, type);
859 return __vhost_get_user_slow(vq, addr, size, type);
862 #define vhost_put_user(vq, x, ptr) \
866 ret = __put_user(x, ptr); \
868 __typeof__(ptr) to = \
869 (__typeof__(ptr)) __vhost_get_user(vq, ptr, \
870 sizeof(*ptr), VHOST_ADDR_USED); \
872 ret = __put_user(x, to); \
879 #define vhost_get_user(vq, x, ptr, type) \
883 ret = __get_user(x, ptr); \
885 __typeof__(ptr) from = \
886 (__typeof__(ptr)) __vhost_get_user(vq, ptr, \
890 ret = __get_user(x, from); \
897 #define vhost_get_avail(vq, x, ptr) \
898 vhost_get_user(vq, x, ptr, VHOST_ADDR_AVAIL)
900 #define vhost_get_used(vq, x, ptr) \
901 vhost_get_user(vq, x, ptr, VHOST_ADDR_USED)
903 static void vhost_dev_lock_vqs(struct vhost_dev *d)
906 for (i = 0; i < d->nvqs; ++i)
907 mutex_lock_nested(&d->vqs[i]->mutex, i);
910 static void vhost_dev_unlock_vqs(struct vhost_dev *d)
913 for (i = 0; i < d->nvqs; ++i)
914 mutex_unlock(&d->vqs[i]->mutex);
917 static int vhost_new_umem_range(struct vhost_umem *umem,
918 u64 start, u64 size, u64 end,
919 u64 userspace_addr, int perm)
921 struct vhost_umem_node *tmp, *node = kmalloc(sizeof(*node), GFP_ATOMIC);
926 if (umem->numem == max_iotlb_entries) {
927 tmp = list_first_entry(&umem->umem_list, typeof(*tmp), link);
928 vhost_umem_free(umem, tmp);
934 node->userspace_addr = userspace_addr;
936 INIT_LIST_HEAD(&node->link);
937 list_add_tail(&node->link, &umem->umem_list);
938 vhost_umem_interval_tree_insert(node, &umem->umem_tree);
944 static void vhost_del_umem_range(struct vhost_umem *umem,
947 struct vhost_umem_node *node;
949 while ((node = vhost_umem_interval_tree_iter_first(&umem->umem_tree,
951 vhost_umem_free(umem, node);
954 static void vhost_iotlb_notify_vq(struct vhost_dev *d,
955 struct vhost_iotlb_msg *msg)
957 struct vhost_msg_node *node, *n;
959 spin_lock(&d->iotlb_lock);
961 list_for_each_entry_safe(node, n, &d->pending_list, node) {
962 struct vhost_iotlb_msg *vq_msg = &node->msg.iotlb;
963 if (msg->iova <= vq_msg->iova &&
964 msg->iova + msg->size - 1 > vq_msg->iova &&
965 vq_msg->type == VHOST_IOTLB_MISS) {
966 vhost_poll_queue(&node->vq->poll);
967 list_del(&node->node);
972 spin_unlock(&d->iotlb_lock);
975 static int umem_access_ok(u64 uaddr, u64 size, int access)
977 unsigned long a = uaddr;
979 /* Make sure 64 bit math will not overflow. */
980 if (vhost_overflow(uaddr, size))
983 if ((access & VHOST_ACCESS_RO) &&
984 !access_ok(VERIFY_READ, (void __user *)a, size))
986 if ((access & VHOST_ACCESS_WO) &&
987 !access_ok(VERIFY_WRITE, (void __user *)a, size))
992 static int vhost_process_iotlb_msg(struct vhost_dev *dev,
993 struct vhost_iotlb_msg *msg)
997 vhost_dev_lock_vqs(dev);
999 case VHOST_IOTLB_UPDATE:
1004 if (umem_access_ok(msg->uaddr, msg->size, msg->perm)) {
1008 vhost_vq_meta_reset(dev);
1009 if (vhost_new_umem_range(dev->iotlb, msg->iova, msg->size,
1010 msg->iova + msg->size - 1,
1011 msg->uaddr, msg->perm)) {
1015 vhost_iotlb_notify_vq(dev, msg);
1017 case VHOST_IOTLB_INVALIDATE:
1022 vhost_vq_meta_reset(dev);
1023 vhost_del_umem_range(dev->iotlb, msg->iova,
1024 msg->iova + msg->size - 1);
1031 vhost_dev_unlock_vqs(dev);
1034 ssize_t vhost_chr_write_iter(struct vhost_dev *dev,
1035 struct iov_iter *from)
1037 struct vhost_msg_node node;
1038 unsigned size = sizeof(struct vhost_msg);
1042 if (iov_iter_count(from) < size)
1044 ret = copy_from_iter(&node.msg, size, from);
1048 switch (node.msg.type) {
1049 case VHOST_IOTLB_MSG:
1050 err = vhost_process_iotlb_msg(dev, &node.msg.iotlb);
1062 EXPORT_SYMBOL(vhost_chr_write_iter);
1064 __poll_t vhost_chr_poll(struct file *file, struct vhost_dev *dev,
1069 poll_wait(file, &dev->wait, wait);
1071 if (!list_empty(&dev->read_list))
1072 mask |= POLLIN | POLLRDNORM;
1076 EXPORT_SYMBOL(vhost_chr_poll);
1078 ssize_t vhost_chr_read_iter(struct vhost_dev *dev, struct iov_iter *to,
1082 struct vhost_msg_node *node;
1084 unsigned size = sizeof(struct vhost_msg);
1086 if (iov_iter_count(to) < size)
1091 prepare_to_wait(&dev->wait, &wait,
1092 TASK_INTERRUPTIBLE);
1094 node = vhost_dequeue_msg(dev, &dev->read_list);
1101 if (signal_pending(current)) {
1114 finish_wait(&dev->wait, &wait);
1117 ret = copy_to_iter(&node->msg, size, to);
1119 if (ret != size || node->msg.type != VHOST_IOTLB_MISS) {
1124 vhost_enqueue_msg(dev, &dev->pending_list, node);
1129 EXPORT_SYMBOL_GPL(vhost_chr_read_iter);
1131 static int vhost_iotlb_miss(struct vhost_virtqueue *vq, u64 iova, int access)
1133 struct vhost_dev *dev = vq->dev;
1134 struct vhost_msg_node *node;
1135 struct vhost_iotlb_msg *msg;
1137 node = vhost_new_msg(vq, VHOST_IOTLB_MISS);
1141 msg = &node->msg.iotlb;
1142 msg->type = VHOST_IOTLB_MISS;
1146 vhost_enqueue_msg(dev, &dev->read_list, node);
1151 static int vq_access_ok(struct vhost_virtqueue *vq, unsigned int num,
1152 struct vring_desc __user *desc,
1153 struct vring_avail __user *avail,
1154 struct vring_used __user *used)
1157 size_t s = vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX) ? 2 : 0;
1159 return access_ok(VERIFY_READ, desc, num * sizeof *desc) &&
1160 access_ok(VERIFY_READ, avail,
1161 sizeof *avail + num * sizeof *avail->ring + s) &&
1162 access_ok(VERIFY_WRITE, used,
1163 sizeof *used + num * sizeof *used->ring + s);
1166 static void vhost_vq_meta_update(struct vhost_virtqueue *vq,
1167 const struct vhost_umem_node *node,
1170 int access = (type == VHOST_ADDR_USED) ?
1171 VHOST_ACCESS_WO : VHOST_ACCESS_RO;
1173 if (likely(node->perm & access))
1174 vq->meta_iotlb[type] = node;
1177 static int iotlb_access_ok(struct vhost_virtqueue *vq,
1178 int access, u64 addr, u64 len, int type)
1180 const struct vhost_umem_node *node;
1181 struct vhost_umem *umem = vq->iotlb;
1182 u64 s = 0, size, orig_addr = addr, last = addr + len - 1;
1184 if (vhost_vq_meta_fetch(vq, addr, len, type))
1188 node = vhost_umem_interval_tree_iter_first(&umem->umem_tree,
1191 if (node == NULL || node->start > addr) {
1192 vhost_iotlb_miss(vq, addr, access);
1194 } else if (!(node->perm & access)) {
1195 /* Report the possible access violation by
1196 * request another translation from userspace.
1201 size = node->size - addr + node->start;
1203 if (orig_addr == addr && size >= len)
1204 vhost_vq_meta_update(vq, node, type);
1213 int vq_iotlb_prefetch(struct vhost_virtqueue *vq)
1215 size_t s = vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX) ? 2 : 0;
1216 unsigned int num = vq->num;
1221 return iotlb_access_ok(vq, VHOST_ACCESS_RO, (u64)(uintptr_t)vq->desc,
1222 num * sizeof(*vq->desc), VHOST_ADDR_DESC) &&
1223 iotlb_access_ok(vq, VHOST_ACCESS_RO, (u64)(uintptr_t)vq->avail,
1225 num * sizeof(*vq->avail->ring) + s,
1226 VHOST_ADDR_AVAIL) &&
1227 iotlb_access_ok(vq, VHOST_ACCESS_WO, (u64)(uintptr_t)vq->used,
1229 num * sizeof(*vq->used->ring) + s,
1232 EXPORT_SYMBOL_GPL(vq_iotlb_prefetch);
1234 /* Can we log writes? */
1235 /* Caller should have device mutex but not vq mutex */
1236 int vhost_log_access_ok(struct vhost_dev *dev)
1238 return memory_access_ok(dev, dev->umem, 1);
1240 EXPORT_SYMBOL_GPL(vhost_log_access_ok);
1242 /* Verify access for write logging. */
1243 /* Caller should have vq mutex and device mutex */
1244 static int vq_log_access_ok(struct vhost_virtqueue *vq,
1245 void __user *log_base)
1247 size_t s = vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX) ? 2 : 0;
1249 return vq_memory_access_ok(log_base, vq->umem,
1250 vhost_has_feature(vq, VHOST_F_LOG_ALL)) &&
1251 (!vq->log_used || log_access_ok(log_base, vq->log_addr,
1253 vq->num * sizeof *vq->used->ring + s));
1256 /* Can we start vq? */
1257 /* Caller should have vq mutex and device mutex */
1258 int vhost_vq_access_ok(struct vhost_virtqueue *vq)
1261 /* When device IOTLB was used, the access validation
1262 * will be validated during prefetching.
1266 return vq_access_ok(vq, vq->num, vq->desc, vq->avail, vq->used) &&
1267 vq_log_access_ok(vq, vq->log_base);
1269 EXPORT_SYMBOL_GPL(vhost_vq_access_ok);
1271 static struct vhost_umem *vhost_umem_alloc(void)
1273 struct vhost_umem *umem = kvzalloc(sizeof(*umem), GFP_KERNEL);
1278 umem->umem_tree = RB_ROOT_CACHED;
1280 INIT_LIST_HEAD(&umem->umem_list);
1285 static long vhost_set_memory(struct vhost_dev *d, struct vhost_memory __user *m)
1287 struct vhost_memory mem, *newmem;
1288 struct vhost_memory_region *region;
1289 struct vhost_umem *newumem, *oldumem;
1290 unsigned long size = offsetof(struct vhost_memory, regions);
1293 if (copy_from_user(&mem, m, size))
1297 if (mem.nregions > max_mem_regions)
1299 newmem = kvzalloc(size + mem.nregions * sizeof(*m->regions), GFP_KERNEL);
1303 memcpy(newmem, &mem, size);
1304 if (copy_from_user(newmem->regions, m->regions,
1305 mem.nregions * sizeof *m->regions)) {
1310 newumem = vhost_umem_alloc();
1316 for (region = newmem->regions;
1317 region < newmem->regions + mem.nregions;
1319 if (vhost_new_umem_range(newumem,
1320 region->guest_phys_addr,
1321 region->memory_size,
1322 region->guest_phys_addr +
1323 region->memory_size - 1,
1324 region->userspace_addr,
1329 if (!memory_access_ok(d, newumem, 0))
1335 /* All memory accesses are done under some VQ mutex. */
1336 for (i = 0; i < d->nvqs; ++i) {
1337 mutex_lock(&d->vqs[i]->mutex);
1338 d->vqs[i]->umem = newumem;
1339 mutex_unlock(&d->vqs[i]->mutex);
1343 vhost_umem_clean(oldumem);
1347 vhost_umem_clean(newumem);
1352 long vhost_vring_ioctl(struct vhost_dev *d, int ioctl, void __user *argp)
1354 struct file *eventfp, *filep = NULL;
1355 bool pollstart = false, pollstop = false;
1356 struct eventfd_ctx *ctx = NULL;
1357 u32 __user *idxp = argp;
1358 struct vhost_virtqueue *vq;
1359 struct vhost_vring_state s;
1360 struct vhost_vring_file f;
1361 struct vhost_vring_addr a;
1365 r = get_user(idx, idxp);
1373 mutex_lock(&vq->mutex);
1376 case VHOST_SET_VRING_NUM:
1377 /* Resizing ring with an active backend?
1378 * You don't want to do that. */
1379 if (vq->private_data) {
1383 if (copy_from_user(&s, argp, sizeof s)) {
1387 if (!s.num || s.num > 0xffff || (s.num & (s.num - 1))) {
1393 case VHOST_SET_VRING_BASE:
1394 /* Moving base with an active backend?
1395 * You don't want to do that. */
1396 if (vq->private_data) {
1400 if (copy_from_user(&s, argp, sizeof s)) {
1404 if (s.num > 0xffff) {
1408 vq->last_avail_idx = s.num;
1409 /* Forget the cached index value. */
1410 vq->avail_idx = vq->last_avail_idx;
1412 case VHOST_GET_VRING_BASE:
1414 s.num = vq->last_avail_idx;
1415 if (copy_to_user(argp, &s, sizeof s))
1418 case VHOST_SET_VRING_ADDR:
1419 if (copy_from_user(&a, argp, sizeof a)) {
1423 if (a.flags & ~(0x1 << VHOST_VRING_F_LOG)) {
1427 /* For 32bit, verify that the top 32bits of the user
1428 data are set to zero. */
1429 if ((u64)(unsigned long)a.desc_user_addr != a.desc_user_addr ||
1430 (u64)(unsigned long)a.used_user_addr != a.used_user_addr ||
1431 (u64)(unsigned long)a.avail_user_addr != a.avail_user_addr) {
1436 /* Make sure it's safe to cast pointers to vring types. */
1437 BUILD_BUG_ON(__alignof__ *vq->avail > VRING_AVAIL_ALIGN_SIZE);
1438 BUILD_BUG_ON(__alignof__ *vq->used > VRING_USED_ALIGN_SIZE);
1439 if ((a.avail_user_addr & (VRING_AVAIL_ALIGN_SIZE - 1)) ||
1440 (a.used_user_addr & (VRING_USED_ALIGN_SIZE - 1)) ||
1441 (a.log_guest_addr & (VRING_USED_ALIGN_SIZE - 1))) {
1446 /* We only verify access here if backend is configured.
1447 * If it is not, we don't as size might not have been setup.
1448 * We will verify when backend is configured. */
1449 if (vq->private_data) {
1450 if (!vq_access_ok(vq, vq->num,
1451 (void __user *)(unsigned long)a.desc_user_addr,
1452 (void __user *)(unsigned long)a.avail_user_addr,
1453 (void __user *)(unsigned long)a.used_user_addr)) {
1458 /* Also validate log access for used ring if enabled. */
1459 if ((a.flags & (0x1 << VHOST_VRING_F_LOG)) &&
1460 !log_access_ok(vq->log_base, a.log_guest_addr,
1462 vq->num * sizeof *vq->used->ring)) {
1468 vq->log_used = !!(a.flags & (0x1 << VHOST_VRING_F_LOG));
1469 vq->desc = (void __user *)(unsigned long)a.desc_user_addr;
1470 vq->avail = (void __user *)(unsigned long)a.avail_user_addr;
1471 vq->log_addr = a.log_guest_addr;
1472 vq->used = (void __user *)(unsigned long)a.used_user_addr;
1474 case VHOST_SET_VRING_KICK:
1475 if (copy_from_user(&f, argp, sizeof f)) {
1479 eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd);
1480 if (IS_ERR(eventfp)) {
1481 r = PTR_ERR(eventfp);
1484 if (eventfp != vq->kick) {
1485 pollstop = (filep = vq->kick) != NULL;
1486 pollstart = (vq->kick = eventfp) != NULL;
1490 case VHOST_SET_VRING_CALL:
1491 if (copy_from_user(&f, argp, sizeof f)) {
1495 eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd);
1496 if (IS_ERR(eventfp)) {
1497 r = PTR_ERR(eventfp);
1500 if (eventfp != vq->call) {
1504 vq->call_ctx = eventfp ?
1505 eventfd_ctx_fileget(eventfp) : NULL;
1509 case VHOST_SET_VRING_ERR:
1510 if (copy_from_user(&f, argp, sizeof f)) {
1514 eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd);
1515 if (IS_ERR(eventfp)) {
1516 r = PTR_ERR(eventfp);
1519 if (eventfp != vq->error) {
1521 vq->error = eventfp;
1522 ctx = vq->error_ctx;
1523 vq->error_ctx = eventfp ?
1524 eventfd_ctx_fileget(eventfp) : NULL;
1528 case VHOST_SET_VRING_ENDIAN:
1529 r = vhost_set_vring_endian(vq, argp);
1531 case VHOST_GET_VRING_ENDIAN:
1532 r = vhost_get_vring_endian(vq, idx, argp);
1534 case VHOST_SET_VRING_BUSYLOOP_TIMEOUT:
1535 if (copy_from_user(&s, argp, sizeof(s))) {
1539 vq->busyloop_timeout = s.num;
1541 case VHOST_GET_VRING_BUSYLOOP_TIMEOUT:
1543 s.num = vq->busyloop_timeout;
1544 if (copy_to_user(argp, &s, sizeof(s)))
1551 if (pollstop && vq->handle_kick)
1552 vhost_poll_stop(&vq->poll);
1555 eventfd_ctx_put(ctx);
1559 if (pollstart && vq->handle_kick)
1560 r = vhost_poll_start(&vq->poll, vq->kick);
1562 mutex_unlock(&vq->mutex);
1564 if (pollstop && vq->handle_kick)
1565 vhost_poll_flush(&vq->poll);
1568 EXPORT_SYMBOL_GPL(vhost_vring_ioctl);
1570 int vhost_init_device_iotlb(struct vhost_dev *d, bool enabled)
1572 struct vhost_umem *niotlb, *oiotlb;
1575 niotlb = vhost_umem_alloc();
1582 for (i = 0; i < d->nvqs; ++i) {
1583 mutex_lock(&d->vqs[i]->mutex);
1584 d->vqs[i]->iotlb = niotlb;
1585 mutex_unlock(&d->vqs[i]->mutex);
1588 vhost_umem_clean(oiotlb);
1592 EXPORT_SYMBOL_GPL(vhost_init_device_iotlb);
1594 /* Caller must have device mutex */
1595 long vhost_dev_ioctl(struct vhost_dev *d, unsigned int ioctl, void __user *argp)
1597 struct file *eventfp, *filep = NULL;
1598 struct eventfd_ctx *ctx = NULL;
1603 /* If you are not the owner, you can become one */
1604 if (ioctl == VHOST_SET_OWNER) {
1605 r = vhost_dev_set_owner(d);
1609 /* You must be the owner to do anything else */
1610 r = vhost_dev_check_owner(d);
1615 case VHOST_SET_MEM_TABLE:
1616 r = vhost_set_memory(d, argp);
1618 case VHOST_SET_LOG_BASE:
1619 if (copy_from_user(&p, argp, sizeof p)) {
1623 if ((u64)(unsigned long)p != p) {
1627 for (i = 0; i < d->nvqs; ++i) {
1628 struct vhost_virtqueue *vq;
1629 void __user *base = (void __user *)(unsigned long)p;
1631 mutex_lock(&vq->mutex);
1632 /* If ring is inactive, will check when it's enabled. */
1633 if (vq->private_data && !vq_log_access_ok(vq, base))
1636 vq->log_base = base;
1637 mutex_unlock(&vq->mutex);
1640 case VHOST_SET_LOG_FD:
1641 r = get_user(fd, (int __user *)argp);
1644 eventfp = fd == -1 ? NULL : eventfd_fget(fd);
1645 if (IS_ERR(eventfp)) {
1646 r = PTR_ERR(eventfp);
1649 if (eventfp != d->log_file) {
1650 filep = d->log_file;
1651 d->log_file = eventfp;
1653 d->log_ctx = eventfp ?
1654 eventfd_ctx_fileget(eventfp) : NULL;
1657 for (i = 0; i < d->nvqs; ++i) {
1658 mutex_lock(&d->vqs[i]->mutex);
1659 d->vqs[i]->log_ctx = d->log_ctx;
1660 mutex_unlock(&d->vqs[i]->mutex);
1663 eventfd_ctx_put(ctx);
1674 EXPORT_SYMBOL_GPL(vhost_dev_ioctl);
1676 /* TODO: This is really inefficient. We need something like get_user()
1677 * (instruction directly accesses the data, with an exception table entry
1678 * returning -EFAULT). See Documentation/x86/exception-tables.txt.
1680 static int set_bit_to_user(int nr, void __user *addr)
1682 unsigned long log = (unsigned long)addr;
1685 int bit = nr + (log % PAGE_SIZE) * 8;
1688 r = get_user_pages_fast(log, 1, 1, &page);
1692 base = kmap_atomic(page);
1694 kunmap_atomic(base);
1695 set_page_dirty_lock(page);
1700 static int log_write(void __user *log_base,
1701 u64 write_address, u64 write_length)
1703 u64 write_page = write_address / VHOST_PAGE_SIZE;
1708 write_length += write_address % VHOST_PAGE_SIZE;
1710 u64 base = (u64)(unsigned long)log_base;
1711 u64 log = base + write_page / 8;
1712 int bit = write_page % 8;
1713 if ((u64)(unsigned long)log != log)
1715 r = set_bit_to_user(bit, (void __user *)(unsigned long)log);
1718 if (write_length <= VHOST_PAGE_SIZE)
1720 write_length -= VHOST_PAGE_SIZE;
1726 int vhost_log_write(struct vhost_virtqueue *vq, struct vhost_log *log,
1727 unsigned int log_num, u64 len)
1731 /* Make sure data written is seen before log. */
1733 for (i = 0; i < log_num; ++i) {
1734 u64 l = min(log[i].len, len);
1735 r = log_write(vq->log_base, log[i].addr, l);
1741 eventfd_signal(vq->log_ctx, 1);
1745 /* Length written exceeds what we have stored. This is a bug. */
1749 EXPORT_SYMBOL_GPL(vhost_log_write);
1751 static int vhost_update_used_flags(struct vhost_virtqueue *vq)
1754 if (vhost_put_user(vq, cpu_to_vhost16(vq, vq->used_flags),
1755 &vq->used->flags) < 0)
1757 if (unlikely(vq->log_used)) {
1758 /* Make sure the flag is seen before log. */
1760 /* Log used flag write. */
1761 used = &vq->used->flags;
1762 log_write(vq->log_base, vq->log_addr +
1763 (used - (void __user *)vq->used),
1764 sizeof vq->used->flags);
1766 eventfd_signal(vq->log_ctx, 1);
1771 static int vhost_update_avail_event(struct vhost_virtqueue *vq, u16 avail_event)
1773 if (vhost_put_user(vq, cpu_to_vhost16(vq, vq->avail_idx),
1774 vhost_avail_event(vq)))
1776 if (unlikely(vq->log_used)) {
1778 /* Make sure the event is seen before log. */
1780 /* Log avail event write */
1781 used = vhost_avail_event(vq);
1782 log_write(vq->log_base, vq->log_addr +
1783 (used - (void __user *)vq->used),
1784 sizeof *vhost_avail_event(vq));
1786 eventfd_signal(vq->log_ctx, 1);
1791 int vhost_vq_init_access(struct vhost_virtqueue *vq)
1793 __virtio16 last_used_idx;
1795 bool is_le = vq->is_le;
1797 if (!vq->private_data)
1800 vhost_init_is_le(vq);
1802 r = vhost_update_used_flags(vq);
1805 vq->signalled_used_valid = false;
1807 !access_ok(VERIFY_READ, &vq->used->idx, sizeof vq->used->idx)) {
1811 r = vhost_get_used(vq, last_used_idx, &vq->used->idx);
1813 vq_err(vq, "Can't access used idx at %p\n",
1817 vq->last_used_idx = vhost16_to_cpu(vq, last_used_idx);
1824 EXPORT_SYMBOL_GPL(vhost_vq_init_access);
1826 static int translate_desc(struct vhost_virtqueue *vq, u64 addr, u32 len,
1827 struct iovec iov[], int iov_size, int access)
1829 const struct vhost_umem_node *node;
1830 struct vhost_dev *dev = vq->dev;
1831 struct vhost_umem *umem = dev->iotlb ? dev->iotlb : dev->umem;
1836 while ((u64)len > s) {
1838 if (unlikely(ret >= iov_size)) {
1843 node = vhost_umem_interval_tree_iter_first(&umem->umem_tree,
1844 addr, addr + len - 1);
1845 if (node == NULL || node->start > addr) {
1846 if (umem != dev->iotlb) {
1852 } else if (!(node->perm & access)) {
1858 size = node->size - addr + node->start;
1859 _iov->iov_len = min((u64)len - s, size);
1860 _iov->iov_base = (void __user *)(unsigned long)
1861 (node->userspace_addr + addr - node->start);
1868 vhost_iotlb_miss(vq, addr, access);
1872 /* Each buffer in the virtqueues is actually a chain of descriptors. This
1873 * function returns the next descriptor in the chain,
1874 * or -1U if we're at the end. */
1875 static unsigned next_desc(struct vhost_virtqueue *vq, struct vring_desc *desc)
1879 /* If this descriptor says it doesn't chain, we're done. */
1880 if (!(desc->flags & cpu_to_vhost16(vq, VRING_DESC_F_NEXT)))
1883 /* Check they're not leading us off end of descriptors. */
1884 next = vhost16_to_cpu(vq, READ_ONCE(desc->next));
1888 static int get_indirect(struct vhost_virtqueue *vq,
1889 struct iovec iov[], unsigned int iov_size,
1890 unsigned int *out_num, unsigned int *in_num,
1891 struct vhost_log *log, unsigned int *log_num,
1892 struct vring_desc *indirect)
1894 struct vring_desc desc;
1895 unsigned int i = 0, count, found = 0;
1896 u32 len = vhost32_to_cpu(vq, indirect->len);
1897 struct iov_iter from;
1901 if (unlikely(len % sizeof desc)) {
1902 vq_err(vq, "Invalid length in indirect descriptor: "
1903 "len 0x%llx not multiple of 0x%zx\n",
1904 (unsigned long long)len,
1909 ret = translate_desc(vq, vhost64_to_cpu(vq, indirect->addr), len, vq->indirect,
1910 UIO_MAXIOV, VHOST_ACCESS_RO);
1911 if (unlikely(ret < 0)) {
1913 vq_err(vq, "Translation failure %d in indirect.\n", ret);
1916 iov_iter_init(&from, READ, vq->indirect, ret, len);
1918 /* We will use the result as an address to read from, so most
1919 * architectures only need a compiler barrier here. */
1920 read_barrier_depends();
1922 count = len / sizeof desc;
1923 /* Buffers are chained via a 16 bit next field, so
1924 * we can have at most 2^16 of these. */
1925 if (unlikely(count > USHRT_MAX + 1)) {
1926 vq_err(vq, "Indirect buffer length too big: %d\n",
1932 unsigned iov_count = *in_num + *out_num;
1933 if (unlikely(++found > count)) {
1934 vq_err(vq, "Loop detected: last one at %u "
1935 "indirect size %u\n",
1939 if (unlikely(!copy_from_iter_full(&desc, sizeof(desc), &from))) {
1940 vq_err(vq, "Failed indirect descriptor: idx %d, %zx\n",
1941 i, (size_t)vhost64_to_cpu(vq, indirect->addr) + i * sizeof desc);
1944 if (unlikely(desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_INDIRECT))) {
1945 vq_err(vq, "Nested indirect descriptor: idx %d, %zx\n",
1946 i, (size_t)vhost64_to_cpu(vq, indirect->addr) + i * sizeof desc);
1950 if (desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_WRITE))
1951 access = VHOST_ACCESS_WO;
1953 access = VHOST_ACCESS_RO;
1955 ret = translate_desc(vq, vhost64_to_cpu(vq, desc.addr),
1956 vhost32_to_cpu(vq, desc.len), iov + iov_count,
1957 iov_size - iov_count, access);
1958 if (unlikely(ret < 0)) {
1960 vq_err(vq, "Translation failure %d indirect idx %d\n",
1964 /* If this is an input descriptor, increment that count. */
1965 if (access == VHOST_ACCESS_WO) {
1967 if (unlikely(log)) {
1968 log[*log_num].addr = vhost64_to_cpu(vq, desc.addr);
1969 log[*log_num].len = vhost32_to_cpu(vq, desc.len);
1973 /* If it's an output descriptor, they're all supposed
1974 * to come before any input descriptors. */
1975 if (unlikely(*in_num)) {
1976 vq_err(vq, "Indirect descriptor "
1977 "has out after in: idx %d\n", i);
1982 } while ((i = next_desc(vq, &desc)) != -1);
1986 /* This looks in the virtqueue and for the first available buffer, and converts
1987 * it to an iovec for convenient access. Since descriptors consist of some
1988 * number of output then some number of input descriptors, it's actually two
1989 * iovecs, but we pack them into one and note how many of each there were.
1991 * This function returns the descriptor number found, or vq->num (which is
1992 * never a valid descriptor number) if none was found. A negative code is
1993 * returned on error. */
1994 int vhost_get_vq_desc(struct vhost_virtqueue *vq,
1995 struct iovec iov[], unsigned int iov_size,
1996 unsigned int *out_num, unsigned int *in_num,
1997 struct vhost_log *log, unsigned int *log_num)
1999 struct vring_desc desc;
2000 unsigned int i, head, found = 0;
2002 __virtio16 avail_idx;
2003 __virtio16 ring_head;
2006 /* Check it isn't doing very strange things with descriptor numbers. */
2007 last_avail_idx = vq->last_avail_idx;
2009 if (vq->avail_idx == vq->last_avail_idx) {
2010 if (unlikely(vhost_get_avail(vq, avail_idx, &vq->avail->idx))) {
2011 vq_err(vq, "Failed to access avail idx at %p\n",
2015 vq->avail_idx = vhost16_to_cpu(vq, avail_idx);
2017 if (unlikely((u16)(vq->avail_idx - last_avail_idx) > vq->num)) {
2018 vq_err(vq, "Guest moved used index from %u to %u",
2019 last_avail_idx, vq->avail_idx);
2023 /* If there's nothing new since last we looked, return
2026 if (vq->avail_idx == last_avail_idx)
2029 /* Only get avail ring entries after they have been
2035 /* Grab the next descriptor number they're advertising, and increment
2036 * the index we've seen. */
2037 if (unlikely(vhost_get_avail(vq, ring_head,
2038 &vq->avail->ring[last_avail_idx & (vq->num - 1)]))) {
2039 vq_err(vq, "Failed to read head: idx %d address %p\n",
2041 &vq->avail->ring[last_avail_idx % vq->num]);
2045 head = vhost16_to_cpu(vq, ring_head);
2047 /* If their number is silly, that's an error. */
2048 if (unlikely(head >= vq->num)) {
2049 vq_err(vq, "Guest says index %u > %u is available",
2054 /* When we start there are none of either input nor output. */
2055 *out_num = *in_num = 0;
2061 unsigned iov_count = *in_num + *out_num;
2062 if (unlikely(i >= vq->num)) {
2063 vq_err(vq, "Desc index is %u > %u, head = %u",
2067 if (unlikely(++found > vq->num)) {
2068 vq_err(vq, "Loop detected: last one at %u "
2069 "vq size %u head %u\n",
2073 ret = vhost_copy_from_user(vq, &desc, vq->desc + i,
2075 if (unlikely(ret)) {
2076 vq_err(vq, "Failed to get descriptor: idx %d addr %p\n",
2080 if (desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_INDIRECT)) {
2081 ret = get_indirect(vq, iov, iov_size,
2083 log, log_num, &desc);
2084 if (unlikely(ret < 0)) {
2086 vq_err(vq, "Failure detected "
2087 "in indirect descriptor at idx %d\n", i);
2093 if (desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_WRITE))
2094 access = VHOST_ACCESS_WO;
2096 access = VHOST_ACCESS_RO;
2097 ret = translate_desc(vq, vhost64_to_cpu(vq, desc.addr),
2098 vhost32_to_cpu(vq, desc.len), iov + iov_count,
2099 iov_size - iov_count, access);
2100 if (unlikely(ret < 0)) {
2102 vq_err(vq, "Translation failure %d descriptor idx %d\n",
2106 if (access == VHOST_ACCESS_WO) {
2107 /* If this is an input descriptor,
2108 * increment that count. */
2110 if (unlikely(log)) {
2111 log[*log_num].addr = vhost64_to_cpu(vq, desc.addr);
2112 log[*log_num].len = vhost32_to_cpu(vq, desc.len);
2116 /* If it's an output descriptor, they're all supposed
2117 * to come before any input descriptors. */
2118 if (unlikely(*in_num)) {
2119 vq_err(vq, "Descriptor has out after in: "
2125 } while ((i = next_desc(vq, &desc)) != -1);
2127 /* On success, increment avail index. */
2128 vq->last_avail_idx++;
2130 /* Assume notifications from guest are disabled at this point,
2131 * if they aren't we would need to update avail_event index. */
2132 BUG_ON(!(vq->used_flags & VRING_USED_F_NO_NOTIFY));
2135 EXPORT_SYMBOL_GPL(vhost_get_vq_desc);
2137 /* Reverse the effect of vhost_get_vq_desc. Useful for error handling. */
2138 void vhost_discard_vq_desc(struct vhost_virtqueue *vq, int n)
2140 vq->last_avail_idx -= n;
2142 EXPORT_SYMBOL_GPL(vhost_discard_vq_desc);
2144 /* After we've used one of their buffers, we tell them about it. We'll then
2145 * want to notify the guest, using eventfd. */
2146 int vhost_add_used(struct vhost_virtqueue *vq, unsigned int head, int len)
2148 struct vring_used_elem heads = {
2149 cpu_to_vhost32(vq, head),
2150 cpu_to_vhost32(vq, len)
2153 return vhost_add_used_n(vq, &heads, 1);
2155 EXPORT_SYMBOL_GPL(vhost_add_used);
2157 static int __vhost_add_used_n(struct vhost_virtqueue *vq,
2158 struct vring_used_elem *heads,
2161 struct vring_used_elem __user *used;
2165 start = vq->last_used_idx & (vq->num - 1);
2166 used = vq->used->ring + start;
2168 if (vhost_put_user(vq, heads[0].id, &used->id)) {
2169 vq_err(vq, "Failed to write used id");
2172 if (vhost_put_user(vq, heads[0].len, &used->len)) {
2173 vq_err(vq, "Failed to write used len");
2176 } else if (vhost_copy_to_user(vq, used, heads, count * sizeof *used)) {
2177 vq_err(vq, "Failed to write used");
2180 if (unlikely(vq->log_used)) {
2181 /* Make sure data is seen before log. */
2183 /* Log used ring entry write. */
2184 log_write(vq->log_base,
2186 ((void __user *)used - (void __user *)vq->used),
2187 count * sizeof *used);
2189 old = vq->last_used_idx;
2190 new = (vq->last_used_idx += count);
2191 /* If the driver never bothers to signal in a very long while,
2192 * used index might wrap around. If that happens, invalidate
2193 * signalled_used index we stored. TODO: make sure driver
2194 * signals at least once in 2^16 and remove this. */
2195 if (unlikely((u16)(new - vq->signalled_used) < (u16)(new - old)))
2196 vq->signalled_used_valid = false;
2200 /* After we've used one of their buffers, we tell them about it. We'll then
2201 * want to notify the guest, using eventfd. */
2202 int vhost_add_used_n(struct vhost_virtqueue *vq, struct vring_used_elem *heads,
2207 start = vq->last_used_idx & (vq->num - 1);
2208 n = vq->num - start;
2210 r = __vhost_add_used_n(vq, heads, n);
2216 r = __vhost_add_used_n(vq, heads, count);
2218 /* Make sure buffer is written before we update index. */
2220 if (vhost_put_user(vq, cpu_to_vhost16(vq, vq->last_used_idx),
2222 vq_err(vq, "Failed to increment used idx");
2225 if (unlikely(vq->log_used)) {
2226 /* Log used index update. */
2227 log_write(vq->log_base,
2228 vq->log_addr + offsetof(struct vring_used, idx),
2229 sizeof vq->used->idx);
2231 eventfd_signal(vq->log_ctx, 1);
2235 EXPORT_SYMBOL_GPL(vhost_add_used_n);
2237 static bool vhost_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2242 /* Flush out used index updates. This is paired
2243 * with the barrier that the Guest executes when enabling
2247 if (vhost_has_feature(vq, VIRTIO_F_NOTIFY_ON_EMPTY) &&
2248 unlikely(vq->avail_idx == vq->last_avail_idx))
2251 if (!vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX)) {
2253 if (vhost_get_avail(vq, flags, &vq->avail->flags)) {
2254 vq_err(vq, "Failed to get flags");
2257 return !(flags & cpu_to_vhost16(vq, VRING_AVAIL_F_NO_INTERRUPT));
2259 old = vq->signalled_used;
2260 v = vq->signalled_used_valid;
2261 new = vq->signalled_used = vq->last_used_idx;
2262 vq->signalled_used_valid = true;
2267 if (vhost_get_avail(vq, event, vhost_used_event(vq))) {
2268 vq_err(vq, "Failed to get used event idx");
2271 return vring_need_event(vhost16_to_cpu(vq, event), new, old);
2274 /* This actually signals the guest, using eventfd. */
2275 void vhost_signal(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2277 /* Signal the Guest tell them we used something up. */
2278 if (vq->call_ctx && vhost_notify(dev, vq))
2279 eventfd_signal(vq->call_ctx, 1);
2281 EXPORT_SYMBOL_GPL(vhost_signal);
2283 /* And here's the combo meal deal. Supersize me! */
2284 void vhost_add_used_and_signal(struct vhost_dev *dev,
2285 struct vhost_virtqueue *vq,
2286 unsigned int head, int len)
2288 vhost_add_used(vq, head, len);
2289 vhost_signal(dev, vq);
2291 EXPORT_SYMBOL_GPL(vhost_add_used_and_signal);
2293 /* multi-buffer version of vhost_add_used_and_signal */
2294 void vhost_add_used_and_signal_n(struct vhost_dev *dev,
2295 struct vhost_virtqueue *vq,
2296 struct vring_used_elem *heads, unsigned count)
2298 vhost_add_used_n(vq, heads, count);
2299 vhost_signal(dev, vq);
2301 EXPORT_SYMBOL_GPL(vhost_add_used_and_signal_n);
2303 /* return true if we're sure that avaiable ring is empty */
2304 bool vhost_vq_avail_empty(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2306 __virtio16 avail_idx;
2309 if (vq->avail_idx != vq->last_avail_idx)
2312 r = vhost_get_avail(vq, avail_idx, &vq->avail->idx);
2315 vq->avail_idx = vhost16_to_cpu(vq, avail_idx);
2317 return vq->avail_idx == vq->last_avail_idx;
2319 EXPORT_SYMBOL_GPL(vhost_vq_avail_empty);
2321 /* OK, now we need to know about added descriptors. */
2322 bool vhost_enable_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2324 __virtio16 avail_idx;
2327 if (!(vq->used_flags & VRING_USED_F_NO_NOTIFY))
2329 vq->used_flags &= ~VRING_USED_F_NO_NOTIFY;
2330 if (!vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX)) {
2331 r = vhost_update_used_flags(vq);
2333 vq_err(vq, "Failed to enable notification at %p: %d\n",
2334 &vq->used->flags, r);
2338 r = vhost_update_avail_event(vq, vq->avail_idx);
2340 vq_err(vq, "Failed to update avail event index at %p: %d\n",
2341 vhost_avail_event(vq), r);
2345 /* They could have slipped one in as we were doing that: make
2346 * sure it's written, then check again. */
2348 r = vhost_get_avail(vq, avail_idx, &vq->avail->idx);
2350 vq_err(vq, "Failed to check avail idx at %p: %d\n",
2351 &vq->avail->idx, r);
2355 return vhost16_to_cpu(vq, avail_idx) != vq->avail_idx;
2357 EXPORT_SYMBOL_GPL(vhost_enable_notify);
2359 /* We don't need to be notified again. */
2360 void vhost_disable_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2364 if (vq->used_flags & VRING_USED_F_NO_NOTIFY)
2366 vq->used_flags |= VRING_USED_F_NO_NOTIFY;
2367 if (!vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX)) {
2368 r = vhost_update_used_flags(vq);
2370 vq_err(vq, "Failed to enable notification at %p: %d\n",
2371 &vq->used->flags, r);
2374 EXPORT_SYMBOL_GPL(vhost_disable_notify);
2376 /* Create a new message. */
2377 struct vhost_msg_node *vhost_new_msg(struct vhost_virtqueue *vq, int type)
2379 struct vhost_msg_node *node = kmalloc(sizeof *node, GFP_KERNEL);
2383 node->msg.type = type;
2386 EXPORT_SYMBOL_GPL(vhost_new_msg);
2388 void vhost_enqueue_msg(struct vhost_dev *dev, struct list_head *head,
2389 struct vhost_msg_node *node)
2391 spin_lock(&dev->iotlb_lock);
2392 list_add_tail(&node->node, head);
2393 spin_unlock(&dev->iotlb_lock);
2395 wake_up_interruptible_poll(&dev->wait, POLLIN | POLLRDNORM);
2397 EXPORT_SYMBOL_GPL(vhost_enqueue_msg);
2399 struct vhost_msg_node *vhost_dequeue_msg(struct vhost_dev *dev,
2400 struct list_head *head)
2402 struct vhost_msg_node *node = NULL;
2404 spin_lock(&dev->iotlb_lock);
2405 if (!list_empty(head)) {
2406 node = list_first_entry(head, struct vhost_msg_node,
2408 list_del(&node->node);
2410 spin_unlock(&dev->iotlb_lock);
2414 EXPORT_SYMBOL_GPL(vhost_dequeue_msg);
2417 static int __init vhost_init(void)
2422 static void __exit vhost_exit(void)
2426 module_init(vhost_init);
2427 module_exit(vhost_exit);
2429 MODULE_VERSION("0.0.1");
2430 MODULE_LICENSE("GPL v2");
2431 MODULE_AUTHOR("Michael S. Tsirkin");
2432 MODULE_DESCRIPTION("Host kernel accelerator for virtio");